I may be off majorly (it's late and I've been ill all day), but since the
hipersockets are implemented as an emulation, they can run at a speed that
is close to memory access speed, which is generally much faster than your
typical network. In that case, the overhead from the IP header is not at
all an issue anymore, whereas the processing speed to handle incoming
packets may become more significant. If it is indeed a significant factor
when dealing with hipersockets, the bandwidth increase from using a higher
MTU value comes mainly from the reduction in the amount of packets that
get processed by the device driver, kernel code, and user code, both on the
writer and the reader side.
In Lucius' case, where only one side gets changed (the Linux side) I believe
that the likely case is that there is a significant impact by the processing
needed to send packets, which gets reduced by increasing the MTU (as long as
the writing process can keep the pipe full so full size packets can be sent
out at top speed). I do not know enough about z/OS to even guess at why it
performs well without any changes on its own end. My gut feeling would be
that it's networking stack may simply be adapting to the MTU on the
hipersocket, and thus that it benefits from the higher MTU value. But that
is just a guess (and the rest of my note is theoretical conjecture - it may
all be 100% wrong :)).
As far as the throughput dropping down beyond a certain MTU value... It is
not unlikely that there is a buffersize limitation somewhere that imposes a
limit on the actual MTU that can be sustained.
Kris
On Sun, Oct 19, 2003 at 10:01:30PM -0500, Adam Thornton wrote:
> On Sun, 2003-10-19 at 20:18, Lucius, Leland wrote:
> > >
> > > SLES8A --> ZOS1 = ~56KB/sec!!!!!!!!!!!!!!!!!!!!!!!!!!
> > >
> > If anyone is interested, I was able to increase this to around 55MB/sec by
> > changing SLES8A's MTU to 20480. Anything higher and it drops back down to
> > "K/sec". I haven't had a chance to change the ZOS side yet.
>
> Question:
>
> How much does a large MTU actually help, even when everyone supports it?
>
> I usually leave mine at either 1492 or 1500, regardless of the allowable
> interface maximum, because there's often something in the path that
> doesn't understand how to fragment, and it seems like the gain is
> negligible.
>
> Let's take a look. It's my fuzzy memory that IP overhead is 40 bytes
> per packet, and I'm sure that someone will correct me if I'm wrong.
> This means that at ~1500 byte MTU, I'm looking at on the order of 3% (a
> little less, but close enough) overhead. At a 20K MTU, I'm looking at
> 0.5%. This means that even in an ideal world, I'm squeezing at most an
> extra 2.5% out of my network, and that just doesn't seem worth the
> tradeoff to me in any but the most exceptional cases.
>
> Have I Grievously Missed The Point here somewhere?
>
> Adam
--
Never underestimate a Mage with:
- the Intelligence to cast Magic Missile,
- the Constitution to survive the first hit, and
- the Dexterity to run fast enough to avoid being hit a second time.
